全域式數位乳房攝影影像電腦輔助診斷系統之開發

The Development of Computer-Aided Diagnosis system of Full-Field Digital Mammography Image

蕭卜綱(Pu-Kang Hsiao)

指導教授 ： 蘇振隆

中原大學/工學院/生物醫學工程研究所/碩士(2013年)

https://doi.org/10.6840/cycu201301117

摘要

惡性腫瘤在這幾年一直是人們的一大威脅，而在女性族群中乳癌的死亡率更居於惡性腫瘤的第四名，過去研究指出，早期的乳房篩檢能夠有效的降低乳癌的發病率及死亡率，而組織牽扯在乳癌檢測中是重要的檢測指標，但容易被忽略，加上亞洲女性乳房較西方國家密度來的高在系統診斷上會有所差異而被誤判。本研究目的是希望能夠透過醫學影像之處理技術，用於全域式數位乳房攝影的影像偵測組織牽扯組織，並適合使用於密度較高乳房影像之全自動電腦輔助診斷系統。
　　本研究對於組織牽扯的偵測主要為影像前處理、影像增強、乳房可疑區域分割、特徵參數擷取及乳房組織分類。首先影像前處理主要會使用Otsu法找尋最佳閥值，以此閥值對乳房影像進行區域成長初步之去除背景等資訊。在影像增強部分會使用向量收斂濾波器對乳腺組織做強化，接著以細線化方式突顯乳腺組織方向。乳房可疑區域分割部分會先以手動圈選去除胸大肌部分以減少誤判，再以前一步驟之影像尋找乳腺組織交會點，並由整體乳房平均灰階降低其取樣，最後由剩下之交會點周圍250×250pixels作為可疑區域。在特徵參數擷取部分是使用灰階共生矩陣及紋理頻譜，經由統計結果取得16個有效之紋理特徵。乳房組織分類則是使用倒傳遞類神經網路方式來減少可疑區域的誤判，進而找出組織牽扯區域。本研究所使用之影像為31個含有組織牽扯區域之影像，其中16張作為訓練組，另外15作為測試組。
　　在結果方面，未進行胸大肌分割及平均閥值判斷時，其分類之結果當Sensitivity達到0.96時，所得偽陽性為391個，平均每張影像之偽陽性為26.07個，而在經過處理後優化之系統其所得之分類結果在Sensitivity達到0.98時，分類後所得之偽陽性為205個，平均每張影像之偽陽性為13.67個，因此加入去除胸大肌及平均閥值能有效的降影像中之可疑區域數量，進而提升系統之檢測效能。
　　本研究以發展一套乳房影像組織牽扯偵測系統，雖然在結果上準確率並不是這麼理想，但可以大略的圈選出可疑區域，輔助醫生在組織牽扯診斷上之參考，並提供往後對組織牽扯研究上的一個方向。目前由於檔案類型不同，並不能完整結合乳房腫塊自動偵測系統及乳房為鈣化自動偵測系統，未來若能取得更多DICOM影像加以研究，期望能夠開發出一套完整之乳房影像電腦輔助診斷系統。

關鍵字

電腦輔助診斷系統 ； 全域式數位乳房攝影 ； 組織牽扯 ； 向量收斂濾波器 ； 灰階共生矩陣 ； 紋理頻譜 ； 倒傳遞類神經網路

並列摘要

Malignancy has been a threatening disease to human in these years. The mortality rate of breast cancer is the fourth of malignancy in feminine. In previous researches, breast cancer screening in early stage can effectively reduce the morbidity and mortality of breast cancer. Architectural distortion (AD) is one of important index in breast cancer detection, but it can be ignored easily. The diagnosis of Asian woman in breast cancer is likely to be false positives (FP) because the breast’s density is higher than woman in West. The purpose of this study is to detect AD in Full-Field Digital Mammography (FFDM) image through the technology of medical imaging processing, and develop a computer-aided detection (CAD) system that suitable for the use in high density breast.
This study includes five stages for AD detection which are image preprocessing, image enhancement, segmentation of region of interest, capturing the texture features, and classification of breast tissue. The Otsu’s method was used to find the best threshold to remove the background and noise of image by region growing method during preprocessing stage. Convergence Index Filter for Vector Field and thinning method were used to enhancement breast tissue image. Before segmentation of region of interest, manual operation of circling pectoral muscle was used to reduce FP. The next step, we find intersection of breast tissue in the image, and use average intensity as threshold to reduce ROIs of image. Finally, the ROIs captured by obtaining 250×250 pixels area around intersection point. After capturing the texture features, Gray-Level Co-occurrence Matrix (GLCM) and texture spectrum were used to obtain features. According to the statistic results, 16 effective features are obtained. For classification of breast tissue, artificial neural network (ANN) is used to reduce FP, and find the AD in images. In this study, 31 images were collected, where 16 images and 15 images were served as training set and test set, respectively.
The results show that the sensitivity of this system is 0.96, and 391 FP and 26.07 FP/Image for test set without manual operation of circling pectoral muscle. After system optimization which including uncourting pectoral muscle area and using the average intensity of breast image as threshold, the sensitivity of system goes to 0.98, and total FP and FP/Image decrease to 205 and 13.67, respectively. Moreover, add the segmentation of circle pectoral muscle and threshold of breast average intensity can effectively decrease the number of ROIs of the image.
In this study, an Automatic Architectural Distortion Detection system of Mammography Image was developed. Although the accuracy of the result is not pretty well, the suspicious areas can still be roughly circled. This system gives the doctor supported reference in the diagnosis of architectural distortion, and provides a direction for future research on the architectural distortion. Because of the different file types, we cannot complete combination Masses Detection system and Clusters of Micro-calcification Detection system. It can become a complete Computer-Aided Diagnosis system of Full-Field Digital Mammography Image, after getting more DICOM files to test and study in the future.

並列關鍵字

Convergence Index Filter for Vector Field ； Full-Field Digital Mammography (FFDM) ； artificial neural network (ANN) ； computer-aided detection (CAD) ； Gray-Level Co-occurrence Matrix (GLCM) ； architectural distortion ； texture spectrum

參考文獻

[16]黃國禎，全域數位乳房攝影之微鈣化群自動偵測系統，中原大學醫學工程研究所碩士論文，中壢，民國97年

[2]H. Burrell, A. Evans, A. Wilson, and S. Pinder, “False negative breast screening assessment: What lessons we can learn?,” Clin. Radiol., Vol. 56, no. 5, pp. 385–388, 2001.

[3]J. Brown, S. Bryan, and R. Warren, “Mammography screening: An　incremental cost effectiveness analysis of double versus single reading of mammograms, ” Br. Med. J., Vol. 312, no. 7034, pp. 809–812, 1996.

[6]S. Baeg and N. Kehtarnavaz, “Texture based classification of mass abnormalities in mammograms, ” Computer-Based Medical Systems, 2000. CBMS 2000. Proceedings. 13th IEEE symposium

[7]Arthur Burgess, “On the noise variance of a digital mammography system, ”Medical Physics 31 .7, July 2004:1987-1995